The US Commerce Department published an interim final rule on Monday that extends AI chip export licensing to rented compute, closing the route where a data centre in a third country sells remote access to a buyer who could not legally import the hardware. The rule was written by the Bureau of Industry and Security with the Treasury Department and the National Security Council, and reported by TechCrunch the same day it landed.

The trigger is capacity. A licence is now required to export, re-export or transfer AI chips and servers to a data centre anywhere in the world if that facility may supply compute to a company headquartered in one of the embargoed D:5 countries, a group that includes China, Russia, Iran and North Korea and runs to 21 states in total. The threshold is 12 million total processing performance, an export-control metric that works out at roughly the combined power of 750 Nvidia H100 or H200 accelerators.

What counts as access

The rule covers indirect routes as well as direct sales. Brokers and go-betweens are named, and operators are told to look through their customers' ownership structures to establish who ultimately controls the workload. Tom's Hardware notes that Nvidia had already asked data centre operators this month to keep at least 51 percent of their business with non-Chinese companies, a private arrangement that the rule now backs with a licensing test.

Five countries get their own licence requirement for advanced chips: Malaysia, Singapore, Indonesia, Thailand and Brazil. Commerce called them higher-than-normal diversion risks. Shipments destined for US-headquartered companies in those countries are exempt, which keeps the rule aimed at ownership rather than geography. Facilities already providing covered services have 90 days from the effective date to obtain licences. Commercial recommendation systems, the kind behind shopping and social feeds, sit outside the rule entirely.

The same document withdraws the AI Diffusion Rule published in January 2025, which had tried to cap exports country by country. The current administration said in May that it would discard that framework as a burden on US firms and allies, and Monday's rule replaces it with a control that follows the buyer instead of the border.

The 90 day clock

For operators, the practical work is documentary. A data centre that rents GPU capacity to a mix of startups now has to establish ultimate ownership for each tenant, keep records that survive an audit, and decide whether a licence application is worth the delay on a contract that may run for a few months. That compliance cost lands hardest on the smaller neoclouds that built their business on selling capacity to whoever would pay, and it arrives while memory prices and power contracts are already squeezing margins.

The rule also lands in the middle of the Nvidia leadership crisis, which has made the company's own compliance record a live political subject. Tom's Hardware notes that Blackwell Ultra B300 accelerators were recently found inside a data center used by China Telecom, which is the sort of discovery that moves a rule from drafting to publication. Jensen Huang's resignation two weeks ago put an interim chief executive in charge of a firm that is simultaneously the main supplier, the main investor in its own customers and the main target of the enforcement debate. Reuters reported last week that AI companies in China are racing to lock in overseas compute before enforcement tightens, which is the behaviour the 90 day window invites: contracts signed now, capacity booked now, paperwork filed later. Malaysia has already moved to use power bills as a way of finding GPUs that arrived without the right paperwork, and the two enforcement approaches will meet in the same data halls.